A Smartphone-based Sensing Platform to Model
Aggressive Driving Behaviors
Jin-Hyuk Hong, Ben Margines, Anind K. Dey
Human-Computer Interaction Institute, Carnegie Mellon University
{hjinh7, benjamin.margines}@gmail.com, [email protected]
ABSTRACT
Driving aggressively increases the risk of accidents.
Assessing a person’s driving style is a useful way to guide
aggressive drivers toward having safer driving behaviors. A
number of studies have investigated driving style, but they
often rely on the use of self-reports or simulators, which are
not suitable for the real-time, continuous, automated
assessment and feedback on the road. In order to understand
and model aggressive driving style, we construct an invehicle sensing platform that uses a smartphone instead of
using heavyweight, expensive systems. Utilizing additional
cheap sensors, our sensing platform can collect useful
information about vehicle movement, maneuvering and
steering wheel movement. We use this data and apply
machine learning to build a driver model that evaluates
drivers’ driving styles based on a number of driving-related
features. From a naturalistic data collection from 22 drivers
for 3 weeks, we analyzed the characteristics of drivers who
have an aggressive driving style. Our model classified those
drivers with an accuracy of 90.5% (violation-class) and
81% (questionnaire-class). We describe how, in future work,
our model can be used to provide real-time feedback to
drivers using only their current smartphone.
Author Keywords
Driving assessment, smartphone, in-vehicle sensing
platform
ACM Classification Keywords
H.5.m. Information interfaces and presentation (e.g., HCI):
Miscellaneous.
INTRODUCTION
Driving is an integral part of many peoples’ lives, as it
allows them to travel faster and further to accomplish more.
It even provides for an improved quality of life [29].
However, it also exposes individuals to the risk of being
involved in harmful and even fatal accidents [5, 8]. Not
only are these accidents injurious to the people involved,
but also they are expensive: accident-related costs in 2011
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed for
profit or commercial advantage and that copies bear this notice and the full citation on the
first page. Copyrights for components of this work owned by others than ACM must be
honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on
servers or to redistribute to lists, requires prior specific permission and/or a fee. Request
permissions from [email protected].
CHI 2014, April 26–May 1, 2014, Toronto, Ontario, Canada.
Copyright ©ACM 978-1-4503-2473-1/14/04...$15.00.
http://dx.doi.org/10.1145/2556288.2557321
in the US were estimated to be $299.5 billion [6]. Road
traffic accidents are directly attributable to human behavior
because they are related to vehicle control and sensitive to
the complex driving environment (e.g., weather, traffic) [7,
30]. Significant research has been performed to understand
the relationship between human behavior and road safety,
for example, asking which behavioral factors can indicate
‘good’ or ‘bad’ driving, or which behavioral factors might
convey the risk that a driver poses to others [3,18].
Aggressive driving, a particular type of driving style, has
long been studied due to its strong correlation with
accidents and traffic safety hazards: by one estimate, it was
influential in causing the majority of accidents in the United
States from 2003 to 2007. [1]. This style of driving refers to
a behavior that deliberately or likely increases the risk of
collision and is motivated by impatience, annoyance,
hostility and/or an attempt to save time [35]. A similar
definition by the NHTSA refers to aggressive driving as
that which “directly affects other road users by placing
them in unnecessary danger” [9]. More specifically, it may
involve a variety of behaviors such as excessive speeding,
tailgating, weaving in and out of traffic, and improper lane
changes [17, 35]. Aggressive driving is a major contributing
factor to road accidents and is now perceived as being one
of the most significant problems of modern day driving [33].
By understanding an individual’s driving style and
providing feedback, he or she may change driving styles to
reduce aggressive driving. For example, a monitoring
system may reduce aggressive driving through voluntary
recommendations (e.g., safety belt use [28]) or involuntary
controls (e.g., speed governor [15]). Drivers can also
manage situational and environmental factors that may
motivate or stimulate them to drive more aggressively. To
support this feedback, we need an automated method for
assessing driving style in real-time.
Many approaches have been used to understand and assess
one’s driving style, and in-vehicle sensing technology,
which yields more objective measures of driving style, has
shown its potential for real-time, continuous, automated
assessment and feedback on the road [5, 18, 36]. From
cameras to an OBD2 connector (on-board diagnostic
system that collects variables from the car computer, such
as engine RPM and vehicle speed), a number of in-vehicle
sensing platforms have been developed to collect a variety
of information on the driver, the vehicle and their driving
environment [2, 14]. However, because the systems are
typically cost thousands of dollars, most consumers are not
able to afford to equip their cars, limiting the overall
applicability to the general public. We are interested in a
solution that is accessible to all drivers.
To meet this need, we propose to model aggressive driving
by: 1) constructing a lightweight in-vehicle sensing
platform based on drivers’ own smartphones to provide an
objective measurement of driving style, 2) using a machine
learning (ML) technique with a number of driving-related
features for an automated assessment of driving style, and
3) evaluating our approach with real-life daily driving data.
In the future, such models can be used to identify
aggressive driving in real time and to provide feedback to
the driver to encourage a change in driving style.
Using our lightweight sensor platform, we collected
naturalistic driving data from 22 participants for three
weeks and were able to differentiate between aggressive
and non-aggressive driving styles. To assess driving style,
driver models are built using a machine learning technique
and a number of sensor-based driving features. These
models identify drivers who have an aggressive driving
style with an accuracy of 90.5% (violation-class) and 81%
(questionnaire-class). We discuss the limitations of our
system, describe how it can support driving feedback, and
identify opportunities for future work.
ASSESSMENT OF DRIVING STYLE
The ability to provide feedback to drivers about their
driving style, especially if it is aggressive, can help drivers
to become more aware of and improve their behavior. This
would likely reduce the probability of poor driving
outcomes. Because of this potential, a number of different
assessment approaches have been investigated, including
questionnaires [32], simulated driving environments [20],
in-vehicle sensing technology [31], and expert-based
examinations [26]. Figure 1 ranks these assessment
methods along the dimensions of cost and how easily they
can be applied to evaluations of real-world driving.
Relevant questionnaires have been created that ask about a
driver’s personality and driving styles to understand his risk
of having an accident [4, 7, 33, 38]. The questionnaires are
simple and inexpensive to administer. Those targeted
toward assessing aggressive driving include the Aggressive
Driving Behavior Scale [17], the Driving Vengeance
Questionnaire [37], the Driving Anger Scale [11], and the
Driving Anger Expression Inventory [12]. These
questionnaires have allowed for great progress in studying
driving style and driver aggression [10]. Of these, the most
popular questionnaire is the Manchester Driving Behavior
Questionnaire (DBQ). The DBQ was initially developed to
investigate whether a questionnaire could distinguish
between driving errors and driving violations; the authors
that three factors could explain 33% of the variance in
aggressive driving - violations, hazardous errors, and
nonhazardous errors [32].
Figure 1. Ranking of existing assessment methods
Despite their clear benefits, questionnaires are not ideal
when applied to the creation of an automated and objective
assessment system. The very nature of the data collection
method precludes real-time evaluation: it would be unsafe
to ask drivers to take a questionnaire while driving.
Furthermore, because questionnaires are answered post hoc,
drivers need to rely on their memory, which can be flawed
due to impression-management and self-deception. Consequently, self-reports may be biased and may not yield
sufficient information to interpret vehicle driving, which is
a complex task composed of continuous and reactive
interactions between the driver, the vehicle and their
environment [5, 34, 36, 40]. Taken together, another
approach to studying driving style is required to obtain
objective, real-time data.
An alternative approach for assessing driving style is to use
a simulated driving environment. Here, a driver controls a
vehicle in a virtual driving environment, providing a safe
and somewhat naturalistic way to study driver behavior.
One key benefit of simulators is that they can be
manipulated to put the drivers in situations which could be
unsafe in a real driving environment, e.g., to simulate an
accident or near-accident. Except for academic research or
training purposes however, simulated driving environments
are used less frequently to provide real-time driving
feedback, as they can be quite expensive and risk
misrepresenting reality.
In many countries, an expert-based examination is
administered as an initial test to judge driving performance.
Here, a certified driving expert evaluates and provides
appropriate feedback in real time about a driver’s skill and
style. It has been widely adopted in road tests of driving
ability. Because no other real-time alternatives exist,
expert-based examination is used by default for licensing
tests. However, providing this type of feedback to drivers is
expensive, which is why it occurs only when someone is
initially learning to drive, or when they explicitly pay an
expert to help improve their driving.
Due to great advances in mobile sensing technology, invehicle data recorders have emerged as a powerful tool to
collect data on driving styles [36]. These platforms now
allow a wide range of information to be recorded to
understand interaction between the driver, the vehicle and
the environment [2]. Various sensing platforms have been
exploited, from the OBD2 and radar to cameras or even
smartphones [22, 39], to collect information and understand
driving style [2, 18, 24]. For example, Boyce and Geller
used a smart car equipped with four cameras to monitor the
driver and the roadway environment of the vehicle, and an
on-line computer to measure information like vehicle
velocity, and investigated relationships between age,
personality and driving style [5]. Miyajima et al. analyzed
drivers’ car-following and pedal operation patterns for
driver identification [24]. Also, the 100-Car Naturalistic
Driving Study used an in-vehicle data recorder, an
accelerometer, multiple cameras, and GPS to capture
information on driver performance and behavior in the
moments leading up to a crash [27]. Malta et al. used the
multimodal signals of brake pedal use and driver speech in
addition to vehicle speed to detect incidents from a realworld driving database [23].
Despite the promise of such an approach, these in-vehicle
sensing platforms tend to be too expensive for the average
consumer to own. To solve this problem, some researchers
have turned to smartphones as a sensor platform, which are
quite powerful, widely available, and are often already
owned by drivers [19, 22, 39]. Smartphone owners now
almost always have the device in their possession, and it can
be an accurate proxy to collect a variety of information about
users [13], even while driving. Johnson and Trivedi tried to
characterize driving style based on smartphone-based sensors
including an accelerometer, gyroscope, magnetometer, GPS,
and camera [19], while You et al. used dual-cameras on
smartphones to detect dangerous driving behavior [39].
Johnson and Trivedi used dynamic time warping to detect the
type of maneuvering based on the movement of a smartphone
mounted on a car in a controlled experiment, while You et al.
have only presented the idea of using the dual-camera
without any experimental validation. Despite their issues,
these studies have shown that a lightweight smartphonebased sensing platform can collect useful information related
to driving style.
SENSOR-BASED EVALUATION OF DRIVING STYLE
Lightweight Sensing Platform using Android Phones
Our In-Vehicle SmartPhone Sensing Platform (IV-SP2) is
depicted in Figure 2. It differs from other sensing platforms
that rely on powerful but heavyweight computers and
expensive sensors [2, 14]. The core part of our system is the
driver’s own Android smartphone mounted on the front
dashboard, which not only collects the location and
movement of the vehicle but also accesses data from two
additional sensors. Current Android smartphones are
equipped with many sensing and communication capabilities,
including dual cameras (front and back), microphones, GPS,
a 3D accelerometer, a 3D gyroscope, a 3D magnetometer, a
proximity sensor, an illumination sensor and a barometer,
and WiFi, 3G and Bluetooth. They are powerful and portable
enough for data collection in vehicles, and can even connect
to other sensors through Bluetooth. Our system exploits most
of these sensing modalities, but we do not include cameras
and microphones due to privacy concerns.
In addition to the Android smartphone, we include two other
modules: a Bluetooth-based on-board diagnostic (OBD2)
reader and an inertial measurement unit (IMU) to support
more diverse sensing modalities related to driving style
(Figure 2). The OBD2 reader is connected to the OBD2 port
of a vehicle and reads a variety of information from the
vehicle, such as vehicle speed, engine RPM, and throttle
position. It can stream the captured data via Bluetooth to the
smartphone. The IMU (in our case, a YEI 3-Space Sensor™
Bluetooth), attached to the upper back of the steering wheel
(toward the dashboard) without impacting its operation, and
it captures the wheel movement with an accelerometer,
gyroscope and compass sensors. It is also communicates via
Bluetooth and can download data in real time. In this study,
we collect all available variables every half second; Table 1
summarizes what our data sensing platform collects.
Figure 2. Overview of the Android-based in-vehicle sensing
platform
Source
Android
Variables
GPS (location, speed, bearing), 3D
acceleration, orientation, compass,
gyroscope, linear acceleration, gravity,
rotation vector, illumination, air pressure
OBD2
Engine coolant temperature, engine load,
engine RPM, throttle position, speed
IMU
3D acceleration, gyroscope, compass
Table 1: Smartphone and sensor variables and costs
Cost
$0,
already
purchased
$25
$15-300
Modeling Driving Style
To use this sensed data to assess driving style, there are a
number of steps to perform: preprocessing the data,
extracting features, and creating a model of driving style.
Preprocessing
Ideally, smartphones and IMUs mounted in different vehicles
would have the same orientation that would not change over
time. In practice, however, vehicles usually have different
shapes of steering wheels and dashboards, and users may
mount them at an arbitrary orientation. In order to build a
general model of driving styles across a population, we
convert sensory signals to have a consistent orientation
across all vehicles. A virtual reorientation method [25] is
used to re-calibrate the 3D measurements, by repeatedly
adjusting the 3D orientation of the smartphone and IMU with
respect to their orientation when the vehicle stops moving.
Also, information from the OBD2 is normalized per vehicle
by simply subtracting baseline readings (car not in motion)
from the in-motion readings.
Feature Extraction
To objectively characterize driving style based on sensor data,
we define a number of driving features from the set of three
sensors that are used in building a machine learning (ML)based model of driving style. Because naturalistic data
collection does not force drivers to drive predefined routes,
the features need to be independent of the context of any
specific trip. For example, trips on highways and city roads
are very different in their characteristics and a model built for
one will not work well on the other. To address this, we
instead define six specific driving situations, which we can
use to compare a driver’s driving style more independently of
her own trips, including 1) START: for five seconds after a
start, 2) STOP: for five seconds before a stop, 3) H-SPEED:
while the vehicle is moving faster than 50 kph, 4) TURN:
when the vehicle is turning (turn is detected based on z-axis
acceleration of IMU) due to the road turning or a turn at an
intersection, 5) B-TURN: for the five seconds just before the
vehicle makes a turn, 6) A-TURN: for five seconds just after
the vehicle makes a turn. In addition, we analyze certain
driving features, such as maximum engine RPM and throttle
position, across all driving trips.
To characterize one’s driving style, we define the driving
features based on the variables of the three sensors as shown
in Table 2. The smartphone, mounted on the car, is used to
detect speed (via GPS), speed change (acceleration), and the
lateral and longitudinal acceleration of the car. The OBD2
reader, installed in the vehicle, records more fine-grained
speed and also pulls some vehicle-specific data, such as
engine RPM, accelerator (throttle) position, and the
derivatives of those variables (note that we only use the three
main variables (speed, engine RPM, and throttle position) out
of a number of potential variables since they are commonly
available from the vehicles of all drivers that participated in
this study). From the IMU mounted on the steering wheel,
we first detect turn events based on the change of z-axis
acceleration, and identify three driving situations, including
TURN, B-TURN, and A-TURN. IV-SP2 produces a sample
composed of the driving features every half second.
Source
Smartphone
Features
Max/avg/std of speed, speed change,
longitudinal/lateral acceleration
OBD2
Max/avg/std of speed, speed change, engine
RPM, engine RPM change, throttle position,
throttle position change
IMU
Used to detect turn events based on its z-axis
acceleration change
Table 2: Features used in characterizing driving style
Aggressive Driving Style Ground Truth
Samples composed of the driving features defined in Table
2 are used to predict a driver’s driving style by evaluating
them for aggressiveness. To acquire the ground truth of
whether a driver is aggressive, we use two methods: selfreports of accidents and ticket records, and a driving style
questionnaire.
Measure 1: Driving violations
As one measure of aggressive driving style, a driver can be
characterized by his record of driving violations. We
characterize a driver as aggressive (violator) or not (nonviolator) using the number of accidents or tickets from the
past 3 years. This time frame ensures that there is sufficient
time for any accidents and tickets to accumulate, but is
short enough so as not to mischaracterize drivers who have
changed their driving patterns. Compared to questionnaire
responses, this measure is more objective and more directly
represents the risk of poor or aggressive driving style
(unless a driver misrepresents his driving record). However,
it lacks the nuance of questionnaires because it does not
distinguish between the causes of the accidents or tickets.
Measure 2: Questionnaire responses
To complement the driving outcome record, we
characterize drivers’ driving style differently based on the
responses to the DBQ [32] and additional aggressive
driving related questions [7, 16]; drivers who score high are
characterized as relatively aggressive and the others as
relatively calm. Incorporating the latter questions that cover
similar, but non-overlapping behaviors of aggressive
violations and ordinary violations [21], allows a clearer
division into aggressive and calm groups by allowing
aggressive drivers to accrue more points. Aggressive
violations are when a driver inconveniences or acts
vengefully toward another driver, e.g., flashing his
headlights for driving too slowly, which clearly falls under
the common denotation for aggressive driving. Ordinary
violations are when a driver violated highway laws and
acted unsafely, but not aggressively towards another driver,
e.g., driving too quickly or not coming to a complete stop,
also refers to unsafe driving practices which can be
improved through feedback. Aggressive violations and
ordinary violations are given the same weight and summed
as a measure of driving aggressiveness.
Modeling Driving Style
To create a machine learning (ML)-based model of driving
style, we build a feature vector using the variables in Table
2 and use it to predict the target driving style. From the
extracted driving features on the 6 driving situations of each
trip of a driver, we first produce a feature vector (called
trip-profile). A new feature vector (called driver-profile) for
each driver is constructed by summarizing a number of his
trip-profiles obtained over 3 weeks. From the driverprofiles and ground truth of drivers, the ML-based driver
model is built to evaluate a driver’s aggressiveness. Given
an input driver-profile feature vector, the driver model
evaluates a driver’s driving style: violator, non-violator,
aggressive, or calm, from our two measures.
Before creating the model, all the continuous driving
features are discretized into five states {very low, low,
medium, high, very high} where the range of values in each
state is chosen to make the number of samples equivalent
across the states. Then to create the driver model, we select
informative driving features based on their information gain
and apply a probabilistic technique, a naive Bayes classifier
(NB), to model the relationship between driving style and
the selected driving features. NB is a simple probabilistic
model based on Bayes' theorem and independence
assumptions of features. Despite its use of over-simplified
assumptions, it often exhibits good performance on many
real-world problems and the resulting model can be easily
interpreted and understood.
In modeling driving style from our driving features, one
goal is to assess driving style with a set of affordable
sensors. To this end, we create three models to predict
driving styles using the driver-profiles.
• Model 1, Smartphone Only: This model does not require
any installation of external sensors. It will capture only the
general movement of a vehicle.
• Model 2, Smartphone and OBD2: This model only
requires the simple and quick installation of an OBD2
reader. The OBD2 reader costs about $25. This model
provides more accurate and detailed information about
vehicle operation.
• Model 3, Smartphone, OBD2, and IMU: The sensors for
this model cost more than the other models because of the
IMU. However, a Nintendo Wii remote is a potential
replacement for our IMU, and costs only $15 and provides
3D acceleration. (Note that our choice of IMU instead of
the Wii remote is due to the usability of its programmatic
interface and usability in a car.) This model additionally
incorporates information about steering wheel control,
when mounted on the steering wheel.
driving violations for the past 3 years, and complete our
driving behavior questionnaire. For each question,
participants ranked the frequency of a driving behavior, on
a six-point Likert-scale (ranging from 0=never, 5=nearly all
the time). Participants were required to visit our lab each
week for data downloading, and were compensated with
$60 for completing 3 weeks of data collection. A total of
542 hours of driving data were collected from 1,017 trips.
RESULT & ANALYSIS
Questionnaire Analysis
According to the self-reports of accidents and speed
violations, we divided participants into two groups: violator
and non-violator (we will call this group division violationclass). As an additional measure of aggressive driving style
based on the questionnaires, we also divided our
participants into two groups: relatively aggressive and
relatively calm. (Note that the term relatively indicates that
our participants cover a subpopulation of drivers, the
division is measured by looking at the relative interquartile
ranges for aggressive versus calm drivers, and we will call
this group division questionnaire-class and use aggressive
and calm without mentioning the term relatively in the rest
of the paper).
Figure 3 shows the basic distribution of our population in
terms of their violation-class and questionnaire-class. Out
of the 22 participants, 12 were characterized as violators;
these drivers had an average of 2 accidents and tickets. 11
of the participants were characterized as aggressive, where
they had an average score of 28.1% (σ=0.05) from the
questionnaire, compared to calm drivers’ scoring 16%
(σ=0.05). While there was a significant participant overlap
between violators and aggressive drivers (10 participants),
the participants’ questionnaire scores were not always
correlated with their accident/ticket record. The correlation
across all drivers is 0.26.
METHOD
We recruited 22 licensed drivers (11 males and 11 females)
from a mid-sized city in the United States, ranging in age
from 21 to 34 years (M = 26). A prescreening questionnaire
was used to recruit a number of people in our target
population: drivers who owned and used an Android
smartphone and drove at least several hours per week.
Participants were asked to drive their own vehicles, which
we equipped with our sensing platform, and to use their
usual daily driving routes plus a predesigned 10-minute
driving course (once per week) to verify the proper
operation of the sensing platform. As described earlier, all
sensor data are collected and stored on their smartphone.
During their first visit to our lab, we installed an IMU and
an OBD2 reader in their vehicles and a logging application
on their smartphones, and provided brief instructions on
how to use the system. Then, they were asked to list their
Figure 3. Distribution of Accident & Ticket Record and
Questionnaire Responses
Sensor-based Driving Style Analysis
In order to have an understanding of the relationship
between the sensor data and driving style, we first
compared the driving features from drivers (10) who were
classified as aggressive violators (i.e., those who were
classified as both aggressive and violators) versus calm
non-violators (6). This allows the data to more clearly
define aggressive driving style. The remaining 6
participants who were either aggressive non-violators or
calm violators will be discussed later in this paper.
Classifying the sensor-based driving feature data as coming
from aggressive violators or calm non-violators yields two
findings. First, in line with intuition, the average
aggressive-violator generally experiences higher g-forces
(forces the body is subjected to during acceleration).
Second, aggressive violators have larger within-person and
between-people variation, which signifies that their
individual driving style is more inconsistent.
aggressive violator accelerates from a stop, she has a
maximum lateral acceleration of .38g versus .35g for calm
non-violators, and a maximum longitudinal acceleration
of .39g versus .31g for calm non-violators. These are the
same g-forces that put strain on the vehicle and can be felt
by passengers in the car.
Aggressive Drivers and Violators Have Higher G-Forces
Contrary to our intuition, both aggressive violators and
calm non-violators have similar maximum speeds, as
shown in Figure 4a; for naturalistic driving data, speed is
not a good way to distinguish aggressive and calm drivers.
However, its derivative, acceleration, proves to be.
Aggressive violators have higher maximum acceleration
and deceleration (derived from speed) throughout their trips
in general and also in specific situations, such as during a
turn (Figure 4b), when driving at high speed (Figure 4c),
accelerating from a stopped position, and decelerating to
stop their vehicles.
Figure 5. Speed comparison on START
Figure 4. Driving feature distribution on aggressive driving style
Furthermore as shown in Figure 5, in the 5 seconds after an
aggressive violator begins moving the car, the car has a
higher maximum speed and engine RPM, but shows the
same average speed and engine RPM as for a calm nonviolator. Figure 5 implies that aggressive violators tend to
wait longer to make a change to their speed, and then make
a more forceful change. In contrast calm non-violators may
respond sooner and make lower magnitude changes.
Many driving features extracted from our sensing platform
more directly support the fact that aggressive violators have
higher g-forces for both acceleration and deceleration, such
as higher smartphone acceleration in the vehicle’s
longitudinal (y-axis) and lateral (x-axis) directions. For
instance, the acceleration of the smartphone shows
aggressive violators experience higher g-forces than calm
non-violators when travelling at high speeds (>50 kph),
which means the higher g-forces they experience are not
limited to starting and stopping the vehicle, but also to
highway travel (e.g., a quick deceleration from 100kph to
60kph or weaving between lanes in high speed traffic with
frequent lane changes). More concretely, when an
Figure 6. Consistency analysis of drivers
Aggressive Drivers and Violators Behave Less Consistently
Aggressive violators, in contrast to calm non-violators,
have an inconsistent driving style, as evidenced by several
driving features, such as the change of speed and the
selection of engine RPMs or throttle positions in our
driving situations, extracted from the sensor data. First,
aggressive violators have lower levels of group consistency
in general, meaning that drivers who are classified as
aggressive by the questionnaire have a greater variation for
many metrics than do drivers who are classified as calm
non-violators. This is measured by looking at the relative
interquartile ranges (IQR) for aggressive violators versus
calm non-violators. Aggressive violators have a broader
distribution of maximum longitudinal acceleration for the 5
seconds before the vehicle comes to a stop as shown in
Figure 6. There is a similar result for the maximum engine
RPM for the 5 seconds after the vehicle makes a turn. This
finding implies that one potential application of this
research could be to identify a particular driver behavior as
being aggressive or calm, and then provide feedback to the
driver on the specific behavior that needs to be corrected.
This finding also indicates that an aggressive violator
driving style is characterized by multiple driving features.
Because some aggressive violators have a lower maximum
RPM and longitudinal acceleration than some calm nonviolators, we must use a sophisticated model that
incorporates a number of driving features.
Second, aggressive violators have lower levels of
individual consistency in most of our driving features (as
calculated from their average standard deviations over all of
their trips), meaning that an average aggressive driver will
be on the lower end of a variable for some trips, and on the
higher end of that same spectrum on other trips. In contrast,
an average calm non-violator will behave more similarly
from trip to trip. For instance, as shown in Figure 6b, calm
non-violators have a lower standard deviation of
longitudinal g-forces when coming to a stop. This finding
also holds when looking just at high-speed driving, after
starting the car, and during a turn. Especially when drivers
are travelling faster than 50 kph, the maximum, average and
standard deviation of the lateral g-forces (measured by the
x-axis acceleration of the smartphone) of aggressive
violators is larger than those of calm non-violators. This
indicates that an average aggressive driver subjects the car
and its passengers to a larger variety of g-forces than do
calm non-violators, who are more consistent.
Building a Machine Learning Model
From our feature analysis, we created two driver models
that predict whether a driver has an aggressive driving style
or not, one determined by the record of accidents and
speeding tickets (violator-class), and another determined by
the questionnaire responses (questionnaire-class). To model
the relationship of driving features from the three data
sources (smartphone, OBD2, and IMU) with each class, we
perform feature selection using information gain and use a
naïve Bayes classifier with 5-bin discretization, and we
conduct leave-one subject-out cross validation for
performance evaluation. Figure 7 shows the result.
IV-SP2-Based Model Accuracy
The average model accuracy with all three sensors was
90.5% (2 false negatives) when predicting the violationclass and 81% (2 false negatives & 2 false positives) when
predicting the questionnaire-class. These are particularly
high accuracies, compared with the baseline model (ZeroR,
which always predicts the most frequent classification):
57% and 52.4% for violation-class and questionnaire-class,
respectively. We also found that our driver model
performed relatively well without using all the sensors in
the IV-SP2 system. More specifically, when modeling the
violation-class with just the smartphone and the OBD2
(without considering turn situations (B-TURN, TURN, and
A-TURN, identified by steering wheel information, using
the IMU), we achieved 81% accuracy. With only the
smartphone (without utilizing car information, such as RPM
and throttle position), we achieved 66.7% accuracy.
Similarly, when predicting the questionnaire-class, the
model is 76.2% accurate with the smartphone and OBD2
only, and 66.7% with just the smartphone.
Figure 7. Prediction of aggressive driving style. Model 1 uses
smartphone data only, Model 2 uses smartphone and OBD2
data, and Model 3 uses smartphone, OBD2 and IMU data.
Participant S18, a violator, provides an example of a
misclassification when fewer sensors are used to predict
driving style. When using the 3-sensor IV-SP2 system, S18
is classified correctly as a violator. However, with only the
smartphone and OBD2 reader, he is classified incorrectly as
a non-violator. This occurred because his general driving
style is mild but he is very aggressive in his turning
behavior, e.g., he often accelerates more than an average
violator while turning. As this behavior occurs frequently,
the 3-sensor model correctly characterizes him as a violator.
With our sensor-based driving feature data, we achieved
higher model accuracy when predicting the violation-class
as opposed to the questionnaire-class (90.5% vs. 81%).
DISCUSSION
In the discussion section, we will address issues on the
applicability of our IV-SP2 and driver models in practice,
limitations of using the driving questionnaire, and difficulty
of conducting naturalistic driving data collection.
Applying IV-SP2 to Real-World Driving in Practice
Using our driving models, we can assess driving style at
three different levels. First, our sensing platform can collect
one’s naturalistic driving data and, based on all trips taken,
let the driver know whether she has an aggressive driving
style which may lead to an increased risk of accidents. This
level of assessment can be performed periodically to coach
drivers “at risk” for aggressive driving styles, e.g., retraining drivers who have at-fault accidents or who have
received a number of speed tickets over some time period,
or to monitor for a decrease in driver performance due to
aging or cognitive disorder [3]. Second, the driver models
can evaluate one’s driving style on a particular trip. Due to
a number of internal and external factors, e.g., stress [4],
heavy traffic and road construction, aggressiveness can
increase, which may increase the risk of an accident.
Objective feedback about an elevated risk may encourage
the driver to correct his or her driving. Figure 8 shows the
evaluations of driving style for each trip for each participant.
Drivers characterized as being aggressive have more
aggressive trips in general, but are not aggressive on all
trips. Likewise, a non-aggressive driver sometimes drives
aggressively.
participants classified as aggressive, but also as a nonviolator) or “unlucky” (4 participants classified as calm, but
also as a violator). Especially, being referred to as unlucky
could be because they have been really unlucky, or because
their poor driving is due to errors and lapses (which are not
captured in our questionnaire), or because their responses
on the questionnaires are not accurate.
Second, for the majority of driving features we examined,
the distinction between violators and non-violators was
much clearer than the distinction between aggressive and
calm drivers. However, when comparing aggressive versus
calm drivers, we cannot so easily find a distinction: the
averages and IQRs are much more similar. Thus, when we
group participants according to their more objective
accident and ticket record, the distinction between groups is
much clearer than when- we group them solely by their
self-reported driving behavior.
Figure 8. Assessment of individual trips
Third, our driver models can detect aggressive driving in
specific situations and provide feedback on driving
behaviors. A driver who has an aggressive driving style
could get specific advice on how to change her driving style,
e.g., keeping a constant engine RPM when driving at a high
speed instead of frequently changing speed. Also, if a driver
turns very aggressively, the system can coach him to make
a better/safer turn. By expanding this to a wider variety of
driving situations, it is also possible to provide both more
specific assessments and feedback.
Still, the way that characterizes driving styles has to be
diversified to present a broader spectrum for understanding
aggressive driving styles. For example, considering driving
style as a continuous variable will better reflect reality and
improve the capability of our system in practice.
Limitations of Questionnaire-Based Classification and
its Impact on Our Model
While questionnaires can provide a cost effective and fast
method for assessing driving style, they suffer from a few
weaknesses when they are used as an initial starting point to
build a driver model with machine learning.
Research suggests that behavioral questionnaires yield
responses that are loosely correlated with real world poor
driving outcomes [38]. Assuming that these responses are
less accurate than self-reports of one’s accident and ticket
record, the driver model based on the questionnaire-class
may have inherited the same limitation. Three findings
point us in this direction.
First, one effect of this can be seen by cross referencing the
responses from the questionnaire and the self-reported
accident and ticket record. If the questionnaire responses
were perfectly correlated with real-world poor driving
outcomes, we would expect all aggressive drivers to be
classified as violators, and calm drivers as non-violators. In
fact, for 6 participants (24%), this relationship does not
hold. These drivers could be referred to as “lucky” (2
Third, the sensor-based driver model did not predict the
questionnaire-class as accurately as it did the violationclass (81% vs 90.5%). The misclassifications in the
questionnaire-class model occurred for 3 drivers, which we
believe occurred because these drivers were either lucky or
unlucky. The misclassifications of the model may be due to
the limitation of the more subjective questionnaire
approach. Accidents and tickets are objective records of
driving style or performance, as are the data from the IVSP2 system. Future work in this line of research should
investigate a more accurate ground truth that includes driver
perception and verified accident and ticket data. In addition,
further augmenting with range sensors and cameras will
provide higher-fidelity ground truth and a validation study
using multiple methods would greatly contribute to the field.
Difficulties of Naturalistic Data Collection
Recording naturalistic data provides a more objective view
of driver behavior, but it still presents challenges with data
processing, some of which were easily surmountable, while
others require more work.
For example, because people generally drive on the same
roads over a short period of time, the variety of road type
data for each participant is not the same. While we
attempted to control this by collecting data over 3 weeks,
some calm drivers spent much of their time traveling faster
on the highways, while some aggressive drivers spent a lot
of their driving time sitting at red lights on city streets.
Without controlling where our participants drove, as has
been done in other studies, we could not use those variables
that would be affected by this bias (e.g., speed), when
averaging across all driving behaviors. For the variables, we
were able to find more distinction between aggressive and
calm drivers only in specific driving situations (e.g., high
speed data, or at vehicle starts and stops).
Also, the fact that each participant is driving his or her own
car and using his or her own smartphone creates an
additional source of bias in the data that is difficult to
control for. Some variables coming from the car, such as
the engine load, are very sensitive to the type of car and
transmission in the car. For these variables, we created new
data features based on the difference and performed basic
calibrations such as normalization and subtracting the
baseline. Still, these calibrations are not ideal; further work
is required to improve the quality of the sensor data and
more reliably and accurately extract values of the variables.
Third, one limitation of the ML-based model which only
uses the smartphone is that it uses the GPS data to
determine speed. GPS sampling rates are not high enough
to reliably determine speed with enough nuance; sampling
rates were closer to 1/2 Hz as opposed to 4-5 Hz for the
other sensors. Furthermore, the GPS data becomes
inaccurate around large buildings and in tunnels; for this
reason, even though the phone-only model achieves good
accuracy, adding an OBD2 reader provides an alternate
source of speed information.
Finally, the IV-SP2 is limited in its ability to detect the
driver’s road context, since it is optimized to capture
information on how the driver is controlling the car. We
suspect that variables such as the weather and traffic or
even the body position of the driver, could improve the
classification accuracy by allowing the model to understand
some intentionality behind behaviors. As our system solely
focuses on the driver’s behavior, additional contextual
information could improve the predictability of our model
by allowing a more nuanced look at the causes and
correlates of particular aggressive driving behaviors.
CONCLUSION
In this paper, we introduced the In-Vehicle Smartphonebased Sensing Platform (IV-SP2), which has the potential
to provide real-time feedback to drivers using their own
Android smartphone with relatively cheap sensors. With
this sensing platform, we collected a variety of information
from drivers’ naturalistic driving, e.g., speed, acceleration,
deceleration, engine RPM, throttle position, and steering
wheel movement, and built driver models to predict
aggressive driving style. With naturalistic driving data
collected from 22 drivers, we could identify aggressive
driving (in an accuracy of 90% for violation-class and 81%
for questionnaire-class). In the future, we aim to enhance
IV-SP2 by introducing new driving situations and driving
features to assess more diverse driving styles, and will
develop and deploy an end-user application that provides
three levels of feedback to improve driver’s driving style.
REFERENCES
1. Aggressive driving: research update, American
Automobile Assoc. Foundation for Traffic Safety,
(2009).
2. Amditis, A., Pagle, K., Joshi, S., and Bekiaris, E.
Driver-vehicle-environment monitoring for on-board
driver support systems: lessons learned from design and
implementation. Applied Ergonomics 41, 2 (2010), 225235.
3. Barkley, R.A. and Cox, C. A review of driving risks and
impairments associated with attentiondeficit/hyperactivity disorder and the effects of
stimulant medication on driving performance. J. of
Safety Research 38, 1 (2007), 113–128.
4. Berdoulat, E., Vavassori, D., and Sastre, M. Driving
anger, emotional and instrumental aggressiveness, and
impulsiveness in the prediction of aggressive and
transgressive driving. Accident Analysis and Prevention
50, (2013), 758-767.
5. Boyce, T.E. and Geller, E.S. An instrumented vehicle
assessment of problem behavior and driving style: Do
younger males really take more risks? Accident Analysis
and Prevention 34, 1 (2002), 51-64.
6. Cambridge Systematics. 2011. Crashes vs Congestion:
What’s the Cost to Society? American Automobile
Association, (November 2011). Retrieved Sept. 5, 2013,
7. Chilaoutakis, J.E., Demakakos, P., Tzamalouka, G.,
Bakou, V., Koumaki, M., and Darviri, C. Aggressive
behavior while driving as predictor of self-reported car
crashes. J. of Safety Research 33, 4 (2002), 431-443.
8. Coughlin, J.F., Reimer, B., and Mehler, B. Monitoring,
managing, and motivating driver safety and well-being.
Pervasive Computing IEEE 10, 3 (2011), 14-21.
9. Countermeasures that work: A highway safety
countermeasure guide for state highway safety offices.
National Highway Traffic Safety Administration Fourth
Edition. (2009).
10. Dahlen, E.R., Edwards, B.D., Tubre, T., Zyphur, M.J.,
and Warren, C.R. Taking a look behind the wheel: An
investigation into the personality predictors of
aggressive driving. Accident Analysis & Prevention 45,
(2012), 1-9.
11. Deffenbacher, J.L., Oetting, E.R., and Lynch, R.S.
Development of a driving anger scale. Psychological
Reports 74, 1 (1994), 83-91.
12. Deffenbacher, J.L., Oetting, E.R., and Swaim, R.C. The
Driving Anger Expression Inventory: A measure of how
people express their anger on the road. Behavior
Research and Therapy 40, 6 (2002), 7-737.
13. Dey, A.K., Wac, K., Ferreira, D., Tassini, K., Hong, J.H., and Ramos, J. Getting closer: an empirical
investigation of the proximity of user to their smart
phones. Ubicomp’11, (2011), 163-172.
14. Espinosa, F., Jimenez, J.A., Santiso, E., Gardel, A.,
Perez, D., Casanova, J. and Santos, C. Design and
implementation of a portable electronic system for
vehicle-driver-route activity measurement.
Measurement 44, 2 (2011), 326-337.
15. Evans, L. 1991. Traffic Safety and the Driver. Van
Nostrand, New York, NY.
16. Hennessey, D. and Weisenthal, D.L. Traffic congestion,
driver stress, and driver aggression. Aggressive
Behavior 25, 6 (1999). 409-423
17. Houston, J.M., Farris, P.B., and Norman, M. The
aggressive driving behavior scale: developing a selfreport measure of unsafe driving practices. North
American J. of Psychology 5, 2 (2003), 269-278.
18. Jensen, M., Wagner, J., and Alexander, K. Analysis of
in-vehicle driver behaviour data for improved safety.
Int. J. of Vehicle Safety 5, 3 (2011), 197-212.
19. Johnson, D.A. and Trivedi, M.M. Driving style
recognition using a smartphone as a sensor platform. In
Proc. 11th Int. Conf. Mobile and Ubiquitous Multimedia,
(2011), 1609-1615.
20. Kee, S., Shamsul, B, and Goh, Y Differences of drivers’
driving performance in simulated driving. Research J. of
Applied Sciences 4, 6 (2009), 230-237.
21. Lawton, R., Parker, D., Manstead, A.S.R., and Stradling,
S.G. The role of affect in predicting social behaviors:
the case of road traffic violations. J. of Applied Social
Psychology 27, 14 (1997), 1258-1276.
22. Li, K., Lu, M., Lu, F., Lv, Q., Shang, L., and
Maksimovic, D. Personalized driving behavior
monitoring and analysis and emerging hybrid vehicles.
Pervasive (LNCS 7319), (2012), 1-19.
23. Malta, L., Miyajima, C. and Takeda, K. A study of
driver behavior under potential threats in vehicle traffic.
IEEE Transactions on Intelligent Transportation
Systems 10, 2 (2009), 201-210.
24. Miyajima, C., Nishiwaki, Y., Ozawa, K., Wakita, T.,
Itou, K.,Takeda, K., and Itakura, F. Driver modeling
based on driving behavior and its evaluation in driver
identification. Proc. of the IEEE 95, 2 (2007), 427-437.
25. Mohan, P., PadManabhan, V.N., and Ramjee, R.
Nericell: rich monitoring of road and traffic conditions
using mobile smartphones. SenSys’08, (2008), 323.
26. Mynttinen, S., Sundstrom, A., Vissers, J. Koivukoski,
M., Hakuli, K., and Keskinen, E. Self-assessed driver
competence among novice drivers – a comparison of
driving test candidate assessments and examiner
assessments in a Dutch and Finnish sample. J. of Safety
Research 40, 4 (2009), 301-309.
27. Neale, V.L., Dingus, T.A., Klauer, S., Sudweeks, J. and
Goodman, M. An overview of the 100-car naturalistic
study and findings. ESV Paper 05-0400, (2005).
28. Nichols, J. New opportunities: Can behavior analysis
help with traffic safety? Panel discussion (panel
session). 24th Annual Convention of the Association for
Behavior Analysis, (1998).
29. Owsley, C. Driving mobility, older adults, and quality of
life. Gerontechnology 1, 4 (2002), 220-230.
30. Özkan, T., Lajunen, T., Chliaoutakis, J.E., Parker, D.,
and Summala, H. Cross-cultural differences in driving
behaviours: A comparison of six countries.
Transportation Research Part F: Traffic Psychology
and Behaviour 9, 3 (2006), 227–242.
31. Paefgen, J., Kehr, F., Zhi, Y., and Michaelles, F.
Driving behavior analysis with smartphones: insights
from a controlled field study. In Proc. 11th Int. Conf.
Mobile and Ubiquitous Multimedia, (2012), 36.
32. Reason, J., Manstead, A., Stradling, S., Baxter, J., and
Campbell, K. Errors and violations on the roads: a real
distinction? Ergonomics 33, 10-11 (1990), 1315-1332.
33. Shinar, D. and Compton, R. Aggressive driving: an
observational study of driver, vehicle, and situational
variables, Accident Analysis and Prevention 36, 3
(2004), 429-437.
34. Sundstrӧm, A. Self-assessment of driving skill – A
review from a measurement perspective. Transportation
Research Part F, 11 (2008), 1-9.
35. Tasca, L. A review of the literature on aggressive
driving research. Aggressive Driving Issues Conf.,
(2000).
36. Toledo, T., Musicant, O., Lotan, T. In-vehicle data
recorders for monitoring and feedback on drivers’
behavior. Transportation Research Part C 16, 3 (2008),
320-331.
37. Wiesenthal, D.L., Hennessy, D., and Gibson, P.M. The
Driving Vengeance Questionnaire (DVQ): The
development of a scale to measure deviant drivers’
attitudes. Violence and Victims 15, 2 (2000), 115-136.
38. Winter, J. and Dodou, D. The driver behavior
questionnaire as a predictor of accidents: a metaanalysis. J. of Safety Research 41, 6 (2010), 463-470.
39. You, C.-W., Montes-de-Oca, M., Bao, T.J., Lane, N.D.,
Lu H., Cardone, G., Torresani, L., and Campbell, A.T.
CarSafe: a driver safety app that detects dangerous
driving behavior using dual-cameras on smartphones.
Ubicomp’12, (2012), 671-672.
40. Zhang, Y., Lin, W.C., and Chin, Y.-K. A patternrecognition approach for driving skill characterization.
IEEE Trans. on Intelligent Transportation Systems 11, 4
(2010), 905-916.